# Place an item on an ordered scale

> Place an item on an ordered scale: 15 companies fully do it, 7 with a limit, 30 carry the tag. Every listing in Decision models with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/decision-models.score-rubric/
- Markdown: https://www.anchorterminal.com/niches/decision-models.score-rubric/index.md (~1,750 tokens)
- Slim: https://www.anchorterminal.com/niches/decision-models.score-rubric/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/decision-models.score-rubric/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 15 | 7 | 30 | 8 | 4 |

## What counts as done

A probability per level and an expected score.

Measured, for our own tests: the chosen reading places at least 85 percent of the caller's labelled items in the right level and no critical item below the urgent cut.

How agents ask for it: how urgent is this; what fraud risk level is this transaction; rate this on a 1 to 5 rubric; how severe is this incident

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Celeris-1 Decision (Marqo Inc)**: score returns a probability per level, the probability-weighted average level and a legend, up to 64 levels. Thresholding guidance is generic.
- **Clef (Cloudflare, Inc.)**: score questions over ordered options return a probability per option. No threshold guidance from Cloudflare.
- **Decider (Mark Marosi (Mapika))**: score over 2 to 10 described levels returns a probability per level and the expected level, with a fitted temperature per answer type.
- **Decision 3.0 (vLLM Semantic Router project)**: score questions return a probability per level; the readout temperature is 1.0 with no calibration figure.
- **elephant-1 (Baseten Labs, Inc.)**: score over 2 to 10 ordered levels returns a probability per level and the expected level; a Brier term in training discourages overconfidence.
- **joey-1 (Baseten Labs, Inc.)**: score over 2 to 10 ordered levels returns a probability per level and the expected level; a Brier term in training discourages overconfidence.
- **Liquid d1 (Liquid AI, Inc.)**: score returns the probability-weighted position and a probability per level; the guide's scoring and reranking examples set thresholds on the probabilities.
- **mammoth-1 (Baseten)**: score over 2 to 10 ordered levels returns a probability per level and the expected level; a Brier term in training discourages overconfidence.
- **Microsoft-Decision-1 (Microsoft)**: score over ordered levels returns an expected score and probabilities; Learn says to validate on your own labelled examples and prefer scores for ordering and thresholds over absolute ratings.
- **Drex 1.5 (Nace AI, Inc.)**: score returns the expected score, a legend, a probability per level and confidence; levels are the criteria array in order.
- **OpenAI Decisions API (OpenAI OpCo, LLC)**: score over 2 to 10 ordered levels returns the probability-weighted average level, probabilities and confidence.
- **pplx-decider (Perplexity AI, Inc.)**: score over up to 10 levels returns the probability-weighted score, legend and probabilities; confidence formula is documented.
- **pplx-decider-flash (Perplexity AI, Inc.)**: score over up to 10 levels with expected score, legend and probabilities; same documented confidence formula.
- **ScaleDown (Carbon ScaleDown, Inc.)**: score over 2 to 10 levels returns the probability-weighted expected value and a probability per level, with the calculation worked through in the docs.
- **SPEV (Sprag AI, Inc.)**: score over 2 to 10 levels returns the expected level, probabilities and confidence; vendor figures only.
- **kettei (Sqwish Ltd, trading as Sqwish Labs)**: ordinal decisions over 2 to 255 levels with a rubric line each return probabilities and expected_level.
- **Jev (TypeSafe AI, Inc.)**: score with probabilities per level, expected score and confidence, plus confidence-routing guidance for answers split across levels; fraud triage on 100 transactions and routing on 80 prompts ran against Strands with Jev stronger on fraud.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Laya (Convai Innovations Pvt. Ltd.)**: score questions are supported in one pass at 33 ms, with answer_confidence to gate on, but issue 131 reports laya-multilingual never selecting the first-listed score option and calibration error is high.
- **Drex Max v1 (Nace AI, Inc.)**: score over ordered levels with expected score and probabilities. Vendor figures only; licence cap.
- **Granite Switch 4.1 (International Business Machines Corporation)**: Ordered labels can be scored, but there is no score type, legend or expected level; the caller derives them from label logprobs.
- **Kev (Jared Palmer)**: score over 1 to 255 levels with expected level and probabilities; issue 161 measured a slot-0 logit offset on identical score options, fixed in PR 236, so pin v1.0 or later weights.
- **JEV-27B-VL (AutoTrust AI Lab)**: Rate 0 to 5 on text and images with a probability per level; vendor figures only.
- **RightNow System Zero (RightNow AI, Inc.)**: score returns an expected level, legend and probabilities, flagged uncalibrated.
- **Strands Decider 2B (Amazon Web Services, Inc.)**: score with expected level and probabilities; fraud triage scored 61 of 100 matching words in option descriptions, and the card says confidence bands hold for short classification only.
- **Vela 2.0 (vLLM Semantic Router project (with KR Labs))**: score over 2 to 10 levels returns the expected level and distribution; USAGE.md says Choice and Score confidence are distribution statistics, not calibrated probabilities.

- not this job: GLiClass (Knowledgator Engineering Ltd.): No ordered score type or expected level; labels are unordered and scored independently.

1 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- expected score rounds a critical item into a lower queue when probability is spread
- a one-point rubric cutoff hides borderline cases

## Around it

Before: support.fetch-ticket, payments.fetch-transaction. After: support.set-priority, human-review.

## Where this comes from

Weight 8 of 100 in Decision models: 4 of 41 accounts; the AWS fraud and delivery tests, the Kev score-order issue and the Dataiku score threshold finding

As data: https://www.anchorterminal.com/api/v1/use-cases/decision-models.score-rubric.json, or the MCP tool get_use_case with id decision-models.score-rubric. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
